Data-Mining-Based Detection of Adverse Drug Events Emmanuel - PowerPoint PPT Presentation

Data-Mining-Based Detection of Adverse Drug Events Emmanuel CHAZARD, Cristian PREDA, Béatrice MERLIN, Grégoire FICHEUR, the PSIP consortium, Régis BEUSCART 2009-08-31 1 MIE Sarajevo – August 2009

PSIP: a research project (7 th FP – ICT) Consortium: 13 partners 1/ Hospitals France, Denmark With / without CPOE 2/ Industry : Oracle, IBM, Medasys (CPOE editors) Vidal (pharmaceutical Kbase) 3/ Academic teams Data & Semantic mining, Decision Support Systems, Human Factors Engineering Duration: 40 months (Jan 08  April 2011) 2009-08-31 2 MIE Sarajevo – August 2009

Patient Safety Through Intelligent Procedures in Medication • Adverse Drug events (ADEs) – Altmost 10% of stays – Less than 2% would be declarated – Responsible of 98000 deaths each year in USA • Objective – Propose new methods to prevent ADEs – Develop automated rules to detect them – Integrate rules in a CDS system generating relevant alerts to the physician 2009-08-31 3 MIE Sarajevo – August 2009

Why use data-mining to detect ADEs? • Chart review is time-consuming • With data-mining, we are able to analyze 10,000 records in few minutes • Data-mining may overcome the inevitable limits of expert knowledge : thousands ADEs described in the litterature • Data-mining may detect complex and sometimes combined ADEs that an expert may not necessarily identify 2009-08-31 4 MIE Sarajevo – August 2009

Data-mining based rules generation What we have How we get it Data Data aggregation Events potential causes and effects Statistical analysis (trees…) Statistical associations effects linked to causes Bibliographic analysis Drug linked events = rules Evaluation of rules Confidence of rules In each medical department 2009-08-31 5 MIE Sarajevo – August 2009

Main reasoning to detect ADE • ADE: injury caused by medical management rather than the underlying condition of the patient • The ADE event: – Is not declared in the stay – requires a specific human case review – Is hidden in the data • Data-mining: three-steps procedure – Identify a kind of traceable incident = “effect” – Automatically find a statistical association with some drugs in combination with other causes – The coincidence of causes and effects generate a rule which allow us to detect stays with ADE 2009-08-31 6 MIE Sarajevo – August 2009

Data-mining based rules generation What we have How we get it Data Data aggregation Events potential causes and effects Statistical analysis (trees…) Statistical associations effects linked to causes Bibliographic analysis Drug linked events = rules Evaluation of rules Confidence of rules In each medical department 2009-08-31 7 MIE Sarajevo – August 2009

From available data to causes and effects Data can be considered as: • Cause or context of an ADE • Effect which is a potential manifestation of an ADE Example on diagnosis: • Cause: chronic diseases, reason of the admission • Effect: acute events during the stay Example on lab results: • Cause: abnormality existing at admission • Effect: abnormality got during the stay Example on drugs: • Cause: prescription of the day • Effect: antidotes… 2009-08-31 8 MIE Sarajevo – August 2009

Medical Data Bases 25,000 records • Copenhagen Hospitals (University Hospitals, DK) – Cardiology & internal medicine: 2,700 records • Rouen hospital (University Hospital, FR) – Cardiology & internal medicine: 800 records • Denain hospital (General Hospital, FR) – Surgery: 2,600 records – Gynecology obstetrics: 1,800 records – Medicine A: 1,700 records – Medicine B: 900 records • Lille hospital (University Hospital, FR) – Geriatry: 15,000 records 2009-08-31 9 MIE Sarajevo – August 2009

From available data to events mmol/l Variables have to be natremia 135 interpreted Data : 125 values of natremia min = 135 max = 145 days binary Event : hyponatremia = 1 hyponatremia start on day 2 stop on day 4 hyponatremia = 0 1 elsewhere 0 days 0 4 2 2009-08-31 10 MIE Sarajevo – August 2009

From available data to events Antiepileptic Example on drugs days 5 6 0 1 2 3 4 Data : binary antiepileptic day 4 Antiepileptic antiepileptic day 5 antiepileptic day 6 Event : 1 antiepileptic = 1 start on day 4 stop on day 6 days 0 antiepileptic = 0 0 4 6 before day 4 2009-08-31 11 MIE Sarajevo – August 2009

Data aggregation: overview • Available data: – Complex data scheme with 7 tables, 91 fields – Potentially more than 30000 different variables – Too numerous and redundant codes • E.g. Diagnosis (ICD 10): 18 000 possible codes • E.g. Drugs (ATC) : 5 400 possible codes • => Need to simplify the data • Aggregated data: – One flat table containing one row per stay – 576 “cause or context” variables – 55 “possible effect” variables (lab values++) 2009-08-31 12 MIE Sarajevo – August 2009

Data-mining based rules generation What we have How we get it Data Data aggregation Events potential causes and effects Statistical analysis (trees,…) Statistical associations effects linked to causes Bibliographic analysis Drug linked events = rules Evaluation of rules Confidence of rules In each medical department 2009-08-31 13 MIE Sarajevo – August 2009

Data Mining Methods • General principle of our analysis: 1) investigate the causes associated with the effect 2) Is there any drug among these causes? • Statistical methods already used in PSIP: – Regression Trees (CART) – Multiple Correspondence Analysis – Logistic Regression Analysis – Principal Component Analysis – Association Rules 2009-08-31 14 MIE Sarajevo – August 2009

Appearance of a too low INR (INR < 2) Risk of Thrombosis 1,08% • Ex. of tree obtained from Too high INR at entry? Yes No the method CART (classification and regression tree) Vitamin K antagonist ? Yes Age > 78.5? Yes • We are you looking for No No the variables (causes) 0,8% 29,2% Prokinetic drug ? Yes No associated with an effect 0,5% 7,75% • Can be read at each node: Hypoalbuminemia? Yes • the name of each No 0% variable used in 58,3% Beta lactam antibacterial? the regression No Yes 66,7% • The confidence at 4,8% Age > this level of the 76.25? Yes No 20% 85,7% 2,65% regression 30% 60% 0% 2009-08-31 15 MIE Sarajevo – August 2009

Appearance of a too low INR (patient with anticoagulation) - Rule N ° 1 1,08% Rule enunciation: Too high INR at entry? Yes No Lab(previous too high INR)=1 & MedInfo(age)>78.68 & Lab(previous hypoalbuminemia)=1 ⇒ Appearance of a too low INR Vitamin K antagonist ? Yes Age > 78.5? No Yes No Rule characteristics: 0,8% 29,2% Support: 6 Prokinetic drug ? Yes No Confidence: 86% 0,5% 7,75% 7 stays match the Hypoalbuminemia? conditions, 6 of them Yes No present the effect 0% 58,3% (86%=6/7) Beta lactam antibacterial? No Yes 66,7% Outcomes: 4,8% Age > 0% death 76.25? Yes No 20% 85,7% 2,65% avg duration: 13.4 days 30% 60% 0% 2009-08-31 16 MIE Sarajevo – August 2009

Appearance of a too low INR (patient with anticoagulation) - Rule N ° 2 1,08% Rule enunciation: Too high INR at entry? Yes No Lab(previous too high INR)=0 & Drug(vitamin K antagonist)=1 & Drug(prokinetic)=1 ⇒ Appearance of a too low INR Vitamin K antagonist ? Yes Age > 78.5? Yes No No Rule characteristics: 0,8% 29,2% Support: 4 Prokinetic drug ? Yes No Confidence: 67% 0,5% 7,75% 6 stays match the Hypoalbuminemia? conditions, 4 of them Yes No present the effect 0% 58,3% (67%=4/6) Beta lactam antibacterial? No Yes 66,7% Outcomes: 4,8% Age > 16.67% death 76.25? Yes No 20% 85,7% 2,65% avg duration: 15 days 30% 60% 0% 2009-08-31 17 MIE Sarajevo – August 2009

Data-mining based rules generation What we have How we get it Data Data aggregation Events potential causes and effects Statistical analysis (trees,…) Statistical associations effects linked to causes Bibliographic analysis Drug linked events = rules Evaluation of rules Confidence of rules In each medical department 2009-08-31 18 MIE Sarajevo – August 2009

Appearance of a too low INR (patient with anticoagulation) - Rule N ° 1 1,08% Rule enunciation: Too high INR at entry? Lab(previous too high INR)=1 Yes No & MedInfo(age)>78.68 Too high INR means Hypocoagulation (risk & Lab(previous hypoalbuminemia)=1 of bleeding): INR > 5 ⇒ Appearance of a too low INR Vitamin K antagonist ? Age > 78.5? Interpretation: When a patient is admitted Yes No Yes No Rule characteristics: 0,8% 29,2% for a too high INR (risk of bleeding), if Support: 6 Prokinetic drug ? No Yes Confidence: 86% age>78 and hypoalbuminemia, then a too 0,5% 7,75% low INR (risk of thrombosis) appears with 7 stays match the a 86 % probability. conditions, 6 of them Hypoalbuminemia? No Yes present the effect 0% 58,3% (86%=6/7) Beta lactam antibacterial? No Yes 66,7% Outcomes: 4,8% Age > 76.25? 0% death 20% 85,7% No Yes 2,65% avg duration: 13.4 days 30% 60% 0% 2009-08-31 19 MIE Sarajevo – August 2009